simple_LaneNet

OptimizedLaneNet — Lightweight Real-time Lane Segmentation

Model Overview

• Backbone: Depthwise Separable Convolutions
• Attention: SE Block (reduction=8)
• Residual Blocks: 2 stages
• Decoder: Bilinear upsampling + conv layers
• Input: 270×480 RGB
• Output: 270×480 binary lane mask

Performance

• Latency (End-to-End, pre+infer+post, batch=1): 25.05 ms/image (≈ 39.9 FPS)
• Pure Inference (forward only, batch=1): 1.88 ms/image (≈ 532 FPS)
• Model size: ~1 MB (PyTorch)